Images have a variety of geometric structures, including edges, corners, contours, and textures. Images with different structures can use different transforms to accomplish sparse representation. For more details of the changes, an image’s rich edge information can be divided into blocks, with sparse basis representing different sparse. In this study, researchers propose a sparse representation algorithm based on improved K-Singular Value Decomposition (K-SVD) for image edges, corners, and contours. The improved algorithm breaks through restrictions on the orthogonal basis and uses different orthogonal bases in different feature regions of the image to construct a frame based on the combination of different regions. This paper analyzes the sparse K-SVD algorithm, concluding that the dictionary is more compact, that the sparsity factor is lower, and that it overall has a better effect on sparse image features. The experiments demonstrate that the improved K-SVD algorithm has a better effect on image smoothing, edge contours, and texture features.